Interactive Object Segmentation with Noisy Binary Inputs

Gregory Canal, Sivabalan Manivasagam, Shaoheng Liang, Christopher J. Rozell

Abstract: We consider the problem of interactively specifying an object segment in an image in an efficient and robust manner via binary inputs corrupted by noise. Our method frames interactive segmentation as a communications system with feedback and leverages a simple channel coding scheme to allow a user to select a segment from an ordered lexicon of segments for a given image. We propose an intuitive lexicon based on ellipses (EllipseLex) and evaluate its ability to specify desired object segments over increasing numbers of inputs at various levels of input noise, and compare it to a baseline algorithm. After evaluating the performance of each method on the Microsoft Common Objects in Context (MS-COCO) dataset using several metrics, we find that our method exhibits competitive performance when specifying real-world objects in images.